The inside story of how bitter rivalries and petty fights derailed Google's self-driving supremacy and cost Uber its shot at the crown
- Anthony Levandowski and Chris Urmson helped create the self-driving industry as it exists today.
- The two were polar opposites and fought over issues big and small, complaining about each other to Google founder Larry Page frequently.
- Urmson won the power struggle and Levandowski ultimately decamped for Uber.
- Years of infighting and indecision hampered Google's effort to launch a self-driving product while Uber and others raced to develop it themselves — but not without problems of their own.
- This story is adapted from Insider senior editor Alex Davies' new book, "Driven: The Race to Create the Autonomous Car," out January 5 from Simon & Schuster.
- Visit Business Insider's homepage for more stories.
On January 7, 2016, Anthony Levandowski emailed Larry Page to wish him a happy new year and to pitch him on a new approach to the search giant's self-driving-car research project.
"Chauffeur is broken," Levandowski wrote, using the code name of the effort he had helped launch. "We're losing our tech advantage fast." The Google founder and Alphabet CEO was used to Levandowski's griping. Since he had pushed for control of the team and been rejected a few years earlier, Levandowski had been sidelined, and made a habit of going over his bosses' heads to complain directly to Page. But the engineer had a point. His team had spent seven years and billions of dollars yet had produced nothing close to a commercial product. He and some comrades didn't deny Chris Urmson, who won out over Levandowski for control of the team, was a terrific roboticist. But they thought him too cautious in deploying their tech and blamed him for letting Google's rivals catch up.
In his email, later revealed amid the Waymo v. Uber self-driving lawsuit, Levandowski floated the idea of starting a "Team Mac," a callback to a bit of Silicon Valley lore: In the early 1980s, Apple was working on a new PC to be called the Lisa, but development was slow, the price tag was approaching $10,000, and IBM was dominating the personal-computer market. Steve Jobs wanted to make a much cheaper and superior product. While the main team kept going with the Lisa — which turned out to be a flop — he put a few employees to work on what became the Macintosh, one of the most successful computers of all time. Levandowski, known for his ability to launch ambitious projects with remarkable speed, thought he could do something similar within Google — just not within Chauffeur. The idea went nowhere.
A few weeks later, on January 27, Levandowski emailed Page again, saying "there's just too much BS," with Urmson and other Chauffeur leaders.
"I want to be in the driver seat, not the passenger seat, and right now [it] feels like I'm in the trunk," Levandowski wrote. He was striking out on his own, he said, with a self-driving-truck outfit.
When Urmson heard the news later that day, he showed no hint of hesitation. He marched Levandowski to his desk and had him pack up his things. He then led the 6'6" engineer out of the self-driving office and over to the "public" side of the building and put him in a conference room. Then he called the human-resources department to come deal with the details of the resignation of the man who had over the past dozen years been his competitor, his teammate, his roommate, and his chief rival in a world they had helped create.
A world that started with a $1 million prize in the California desert, that now promised to reinvent a fundamental facet of human life, and to deliver unfathomable riches to those who could make it happen.
The challengers emerge
Levandowski and Urmson first faced off in March 2004, when the Defense Advanced Research Projects Agency hosted the Grand Challenge, a race for autonomous vehicles through California's Mojave Desert. The director of the small and secretive arm of the Pentagon, better known as DARPA, hoped the race, which was open to anyone, would accelerate the development of unmanned vehicles for use in combat and save lives as the US was getting mired in wars in Iraq and Afghanistan.
Urmson came to the Grand Challenge as a leader on the team fielded by Carnegie Mellon University, where he was pursuing a doctorate at perhaps the country's most prestigious robotics institute. CMU's well-funded "Red Team" and its autonomous Humvee, Sandstorm, were the heavy favorites to win the 142-mile race and the $1 million prize.
Levandowski had no such pedigree or support. He was a master's student at the University of California, Berkeley's industrial-engineering department but launched his own effort to build an autonomous motorcycle, called GhostRider. His harebrained creation could stand on its own, but was a dunce when it came to detecting and avoiding obstacles. Levandowski earned a spot in the final round of the race mostly via chutzpah. "It was such a PR attraction that we had to bring it to the desert and see what it did," DARPA's race director said in a 2018 interview.
It ended up failing miserably, flopping to the ground roughly one yard from the starting line. The rest of the field did little better. CMU's Sandstorm went the farthest, getting stuck on a hairpin curve after 7.4 miles.
The Grand Challenge succeeded, though, in whipping up enthusiasm for self-driving technology among amateur and professional engineers across the country. Eighteen months later, at the second Grand Challenge, five vehicles reached the finish line. Urmson's team finished second, losing out to a team led by the head of Stanford's Artificial Intelligence Laboratory, Sebastian Thrun. Levandowski had revamped his motorcycle for the race but fared little better, failing to get past the qualifying round.
DARPA eclipsed that success in 2007 with the Urban Challenge, which swapped the desert roads of the first two races for a mock city at an abandoned Air Force base, complete with intersections, stop signs, and other human-driven vehicles (fitted with protective roll cages).
Here Urmson and Carnegie Mellon finally got their win. Stanford took second, helped along by Levandowski, who had joined the team after impressing Thrun with his scrappy motorcycle effort. But the real upshot of the Urban Challenge was the robotic ballet it presented to the world, with unmanned vehicles nosing through parking lots, turning left into traffic, and acing the annoyingly human four-way stop. It made clear to those watching that self-driving technology was no longer a futuristic fantasy, but a reality not so far out of reach.
Among those watching that day was Larry Page, who as a student had considered studying the tech before taking a professor's advice to focus on internet search. He had reached out to Thrun after the professor won the second Grand Challenge. Page hired him part-time to accelerate the development of Google Street View, using software inspired by his self-driving work. Thrun in turn hired Levandowski. The two engineers and their team mapped a million miles in just seven months.
So when Page decided, near the end of 2008, that he wanted cars that could drive themselves, Thrun was the obvious choice to spearhead the effort.
Thrun called a meeting, at his Lake Tahoe chalet, of the best self-driving engineers he knew inside and outside of Google. Nearly all were veterans of DARPA's challenges, including Urmson and Levandowski. Thrun invited them along for the ride, and they accepted.
Google's vision was grand in its ambiguity. Page wasn't proposing a typical government or defense-industry contract. He wasn't looking for a specific solution to a specific need, with carefully defined funding and spending limits. He just wanted an autonomous car, and he was willing to write something that looked a lot like a blank check to get it. He did, however, believe in aggressive goals. He and Brin pulled up Google Maps and selected 10 routes scattered throughout California, each roughly 100 miles. Thrun, Urmson, Levandowski, and their new teammates would have to make a car that could handle every inch of each route, without human intervention, about 1,000 miles in all. The Chauffeur team, about a dozen strong, called this new challenge the Larry 1K.
Urmson, who had led Carnegie Mellon's Urban Challenge winning team, ran the show from day to day, with Thrun above him. He moved to the Bay Area from Pittsburgh a few months before his wife and two sons did, and while he got settled, he took up Levandowski on an offer to stay in a spare room in his Palo Alto house.
One of the few team members without a Ph.D., Levandowski worked on hardware, setting up a dozen Toyota Priuses with the necessary suite of radars, cameras, and roof-mounted spinning lidar system. As ever, he delivered at remarkable speed. In the DARPA Challenges, teams had typically taken months to get a vehicle up and running. Levandowski put half a dozen cars together in a couple of weeks.
The differences between Levandowski and Urmson were clear from the beginning. The roommates got along well, but they came at the problem — at the world in general — from opposite mind-sets. Levandowski moved as a blur. "If you need to blow up a dam in Germany, Anthony's your guy," Isaac Taylor, Chauffeur's operations manager, said in an interview. "The general who's going to win you World War II."
Urmson was methodical. Near the start of the project, he told Thrun he needed to rewrite the car's core communication software. Thrun, who leaned toward the fast and scrappy side, thought the system they had was fine. Even if it crashed occasionally, stopping to rework it could slow progress. Urmson insisted, and took several weeks to do the work. The result was a stronger, more reliable system. "He was damn right," Thrun later admitted in an interview.
Over the next 18 months, the Chauffeur team improved their system at a remarkable pace, racking up tens of thousands of miles of experience on Bay Area highways, picking off the Larry 1K routes, and breaking up long hours of coding with games of foosball. They completed the last 100-mile stretch in September 2010 and celebrated with a party at Thrun's house, where they threw each other into the pool and played "human foosball" on an inflatable field.
A few weeks later, the New York Times tech reporter John Markoff wrote a story that revealed the secret project to the world. He got a ride in the car and gave a positive review, noting it still struggled with things like understanding traffic cops' hand gestures. He also flicked at a murkier challenge, one that would do much to tear the Chauffeur team apart in the ensuing years, writing: "The company did not yet have a clear plan to create a business from the experiments."
'Make Anthony rich'
The most immediate path to commercializing the technology, it was soon agreed, was to package a version of the software and hardware that automakers could offer buyers as a luxury add-on. On surface streets, where the unpredictable conditions and "edge case" scenarios were still too great for the technology to safely navigate, the car's human owner would drive, just as she normally does. But on the highway she'd have the option to let the car take over, similar to way that a car's cruise-control functions.
It sounded reasonable, provided the Google team could rein in the costs and make the sensors reliable. And so, in late 2010 and early 2011, Urmson and Levandowski began visiting Detroit and Germany to explore partnerships with established automakers.
Far from being intrigued, the auto executives that the two Googlers met with were either dismissive or uninterested. They considered testing on public roads reckless, thought the roof-mounted lidar laser scanner looked impossibly stupid, and said no, they were not interested in working together.
There were problems closer to home, too, as a result of Levandowski's side gigs as the founder of a pair of startups called 510 Systems and Anthony's Robots. Google was buying technology from 510 Systems, initially for Street View, and later for the Chauffeur project.
The unorthodox arrangement meant that Levandowski was effectively sending Google's money across the Bay and into his own pockets, but management had tacitly looked the other way: Google was one of the world's most profitable companies, and Levandowski, as hard-charging as ever, was getting the job done.
By early 2011, however, things were getting awkward. 510 and its sister company, Anthony's Robots, had started developing their own self-driving software and a proprietary lidar scanner. Cock your head just right, and Levandowski looked like a potential competitor to the Google effort he had helped start. What's more, one of 510's biggest customers was apparently hoping to acquire the company, along with all its intellectual property.
Google's chieftains decided that Levandowski's work could not fall into anyone else's hands, and they moved to buy the two companies. But Levandowski's questionable setup raised the eyebrows of David Lawee, the exec who led Google's acquisitions team.
"I don't understand this guy at all," Lawee told Thrun in an email later revealed in court proceedings. "As Google, I suppose I'm prepared to take the risk with Anthony, but I can say definitively that if I was choosing a business partner to start a company with, there is no way in hell that I would proceed."
Levandowski had grown up in Brussels with his mother, a French citizen who worked for the European Union. At 14, he moved to Northern California to live with his father, an American businessman. He wanted to attend high school in the US, he said later, because he thought it would better set him up for professional success. Even if he couldn't write in English when he arrived, he didn't take long to embrace American-style capitalism. At Tamalpais High School, in tony Marin County, Levandowski sold candy to fellow students. After building a website for the school, he launched a business doing the same for companies around town. Before leaving high school, he had made enough money to buy a three-bedroom house (with some help from his dad and stepmother) near the Berkeley campus.
He kept up a web business called La Raison when he started as an undergrad at Berkeley, but realized the only way to keep up with larger competitors was to cut his prices or win over clients with hands-on customer service. "There was no barrier to entry there," he told the university's news site. "I don't want to be someone just providing a commodity at a low price." Levandowski preferred outthinking people to outworking them. His approach to problem-solving was to find the hack, the way to game the system and jump ahead. Like many in Silicon Valley, he saw this not as cheating, but as good engineering.
Read more: Investors are betting these 10 self-driving-car startups are primed to take on Tesla and Waymo
Levandowski was also a fount of charisma. He'd kept his teammates in the DARPA Grand Challenges going with burritos and enthusiasm. In an interview, a colleague at 510 called conversations with Levandowski, who could weave disparate ideas into big-picture imaginings, almost addictive. But he couldn't win everyone over.
Thrun was starting to move away from his leadership of the self-driving team, and had considered putting the tall, young engineer in charge. He soon dropped the idea, replying to Lawee later that day: "I now question whether Anthony is fit to lead Chauffeur. I hear similar concerns from the engineering team. Several people there have contacted me that they have concerns about Anthony's commitment and integrity." Some team members disliked the fact that Levandowski owned a company from which they were buying hardware — it just seemed shady. Levandowski didn't help his reputation by wearing a custom T-shirt reading "I Drink Your Milkshake." It was a reference to the 2007 film "There Will Be Blood," about a conniving oilman who bilks people out of their fortunes, murdering a couple along the way.
"He plays above board," said one Chauffeur teammate in an interview. "He just changes the rules as he goes."
Ultimately, Thrun concluded that putting Levandowski in charge of Chauffeur was too risky. "If he is the single leader, a good number of team members will leave," he told Larry Page in an email made public during the Waymo v. Uber trial. Urmson — reliable and respected — would take the top spot.
The deal to buy Levandowski's technology, and to keep his dynamism in house, went forward, though. Google paid roughly $20 million for 510 Systems and Anthony's Robots. Workers who had joined 510 in its first few years and expected a big payout — standard fare for an acquired startup — were disappointed. Levandowski had sold 510 for just enough so that the real money stayed with the founders; anything above the $20 million mark would have gone to the employees.
The millions that Levandowski reaped from the acquisition were just the amuse-bouche. Chauffeur had an unusual bonus structure: At four-year intervals, Google would come up with a valuation for what they'd built, and pay workers a certain percentage of that number; leaving before the four-year mark meant sacrificing the bonus. Most of the team's key members would get a 0.5% share. But Levandowski, partly to compensate him for selling his companies for a relatively low price, would get a tenth of the entire pot, 20 times more than most of his colleagues. The 10%, according to an email from Thrun, made public during the Waymo v. Uber trial, was based on a directive from Page to "make Anthony rich if Chauffeur succeeds."
That incentive structure made decisions about how Chauffeur would create a business all the more important. As their research moved along, the engineers staked out their own positions not just on who should lead, but on where they should go.
Levandowski, always looking to charge ahead, advocated for selling an aftermarket kit that could make any car drive itself on the highway. Urmson wanted something more carefully integrated into a car, believing it would be more reliable, even if it took longer to produce. The two men were no longer roommates; Urmson had moved out when his wife and sons arrived from Pittsburgh, and had put a down payment on a house with his bonus from completing the Larry 1K. But by now they were finding it hard even to be colleagues.
Where Levandowski had panache, Urmson had pedigree. Born and raised in Canada, he'd graduated from the University of Manitoba with a nearly perfect GPA before heading to Pittsburgh to get his Ph.D. from Carnegie Mellon's illustrious robotics institute. By the time of the 2004 Grand Challenge, he'd helped design robots to trawl for meteorites in Antarctica, assemble space stations in orbit, and search for life on Mars. By 2009, he was an up-and-coming robotics star. He was also, various colleagues at CMU and Google said in interviews, a researcher at heart, more suited to exploring new ideas than to hardening them into products.
True to his Canadian heritage, he was unfailingly polite. He was nice to journalists but never talked out of school. He never bent the rules, or at least was never caught. But some Chauffeur teammates — including Levandowski — said in interviews that Urmson was a cannier political player than he let on. And he was willing to fight, if quietly, for what he wanted.
The pursuit of the Larry 1K had preempted potential tensions by motivating everyone toward a shared, well-defined goal. Now the differences in how they viewed Chauffeur's future, and the path to get there, became a source of conflict, and part of a bigger power struggle. Even after Thrun chose Urmson to lead, Urmson and Levandowski made it clear that they wanted to run things, and that if the other had too much power, the program could be ruined. Thrun thought each offered a valuable skill set, and that differences in perspective were healthy. But this was not Abraham Lincoln's "team of rivals," overcoming their differences to help preserve the Union and win the Civil War. It was more like that war itself, where every topic of discussion could be turned into a debate, an argument, a swear-laden shouting match.
They and their teammates managed to find casus belli in the least important places, like what sorts of buttons the car should have. Some wanted two buttons: green to engage the system, red to disengage. Others thought it would be simpler to have just one for both on and off. They argued over whether they should add new buttons, or repurpose those already in the car. They argued over how to display the speed the driver set for the vehicle, whether as the absolute number (say, 70 mph) or as an offset to the speed limit (65 + 5).
Source: Read Full Article